AI Dev Discovers 88% Accuracy Was Inflated by Train-Test Data Leak

A developer building an AI reasoning system called Origin celebrated hitting 88% accuracy on compositional reasoning probes, only to discover the next morning that 23 of 26 held-out test pairs had already appeared in training data. After fixing the data generators to enforce a strict train-test split and adding a verification tool to block overlapping runs, the honest accuracy dropped to 58% — well below the 60% threshold required to advance to the next development stage. Despite the setback, the real score was still roughly 16 times better than a baseline constant-prediction model. The incident was traced not to a single bug but to lax development discipline, symbolized by files labeled 'sandbox' that were actually being used in production. The developer responded by renaming files and scripts to reflect their true role, aiming to close the gap between informal habits and production-grade standards.